CN112034793A - Method for monitoring production running state space of oxidation working section in PTA process flow - Google Patents

Method for monitoring production running state space of oxidation working section in PTA process flow Download PDF

Info

Publication number
CN112034793A
CN112034793A CN202010793556.6A CN202010793556A CN112034793A CN 112034793 A CN112034793 A CN 112034793A CN 202010793556 A CN202010793556 A CN 202010793556A CN 112034793 A CN112034793 A CN 112034793A
Authority
CN
China
Prior art keywords
data
monitoring
sae
model
dbn
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010793556.6A
Other languages
Chinese (zh)
Other versions
CN112034793B (en
Inventor
颜学峰
于健博
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
East China University of Science and Technology
Original Assignee
East China University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by East China University of Science and Technology filed Critical East China University of Science and Technology
Priority to CN202010793556.6A priority Critical patent/CN112034793B/en
Publication of CN112034793A publication Critical patent/CN112034793A/en
Application granted granted Critical
Publication of CN112034793B publication Critical patent/CN112034793B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
    • G05B19/41885—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by modeling, simulation of the manufacturing system
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/31—From computer integrated manufacturing till monitoring
    • G05B2219/31359—Object oriented model for fault, quality control
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]

Landscapes

  • Engineering & Computer Science (AREA)
  • Manufacturing & Machinery (AREA)
  • General Engineering & Computer Science (AREA)
  • Quality & Reliability (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a method for monitoring the production running state space of an oxidation working section in a PTA process flow, which takes a global variable in the process flow as an input variable and establishes a working state monitoring model to identify the working state of an industrial process at the current moment; the global variable is a state space detection of the content of the byproduct 4-CBA in the production process so as to reflect the working state of the current process flow; the working state monitoring model is established by constructing a deep neural network state monitoring model to extract high-order abstract characteristic information in data; extracting high-order numerical value feature information in the data and high-order correlation features among variables by using a depth self-encoder SAE, and extracting high-order distribution feature information in the data by using a depth confidence network DBN; features of multiple scales are fused to enrich the expression of many aspects of the data. And (3) integrating various characteristics of the data, and monitoring and identifying the state by using a Softmax classifier.

Description

Method for monitoring production running state space of oxidation working section in PTA process flow
Technical Field
The invention belongs to the cross field of petrochemical engineering and process control, and relates to a method for monitoring the production running state space of an oxidation section in the technical process of Purified Terephthalic Acid (PTA).
Background
Purified Terephthalic Acid (PTA) is one of the important bulk organic raw materials, is mainly used for producing polyethylene terephthalate, polytrimethylene terephthalate and polybutylene terephthalate, is also used as a dye intermediate, and is an important material for manufacturing polyester fibers, films and insulating varnish. The fiber is widely applied to various aspects of national economy such as chemical fiber, light industry, electronics, buildings and the like, and is closely related to the living standard of people, living environment and social sustainable development.
The PTA production process has two components: one is xylene (PX) oxidation process; the other is a process for purifying Crude Terephthalic Acid (CTA). The PX oxidation process flow comprises the following steps: para PX is used as a raw material, acetic acid is used as a solvent, tetrabromoethane is used as an accelerant to react with oxygen in the air under the action of cobalt acetate and a manganese acetate catalyst to generate CTA. A large amount of reaction heat evolved in the reaction is taken away by evaporation of the solvent, and this heat is recovered by the by-product steam. And cooling and depressurizing the oxidation reaction liquid by serially connected crystallizers, and filtering and drying to obtain an intermediate product CTA. The main component of the by-product produced in this process is 4-CBA. The CTA refining process flow is as follows: preparing CTA into slurry with a certain concentration by using deionized water, heating to the required dissolving temperature, and then sending to a hydrogenation reactor. The 4-CBA is reduced to PTA by catalytic hydrogenation reaction. Since p-toluic acid is easily soluble in water, p-toluic acid can be separated during recrystallization, separation and drying, and Pure Terephthalic Acid (PTA) with high purity can be obtained.
In the whole production process flow of PTA, the main reaction byproduct is intermediate 4-CBA, generally 4-CBA not only forms eutectic with TA during crystallization to pollute TA, but also influences the polymerization reaction of TA in downstream product production. The high and low content of 4-CBA is an important index of PTA quality, and the main task of the hydrofining process is to reduce the content of 4-CBA in CTA to obtain refined terephthalic acid. CTA is prepared into slurry with a certain concentration by deionized water, heated to the required dissolving temperature and then sent to a hydrogenation reactor. The 4-CBA contained in the crude terephthalic acid is converted into water-soluble substances through catalytic hydrogenation reaction. The hydrogenation reaction liquid is sent to a centrifuge for separation after being cooled and depressurized step by step in serially connected crystallizers, the obtained filter cake is pulped by deionized water, and then is filtered and dried to prepare fiber-grade purified terephthalic acid, wherein the PTA production process flow is shown in figure 1. Wherein, the byproduct 4-CBA is in an extremely important position in the whole production process, and the output content of the byproduct directly reflects the working state of the production process: over oxidation, moderate oxidation and under oxidation, the current working state can be accurately judged based on the over oxidation, the resources can be reasonably distributed, and the product quality of PTA is improved. Therefore, establishing a state monitoring model taking 4-CBA as an index in the PTA production process plays a very important role in optimizing the industrial process.
In recent years, as the production process flow is developed toward the direction of complication and intellectualization, and the sensor technology and the Distributed Control System (DCS) are widely used in the industrial process, the process state monitoring and recognition technology based on data is widely used in this field. The method based on data driving does not need to consider mechanism knowledge in a complex industrial process, only needs to construct a monitoring and identifying model based on data characteristics, has strong universality and convenience, is widely concerned by academia and industry, and is more suitable for modern complex petrochemical production processes. Among the commonly used methods are Principal Component Analysis (PCA), Partial Least Squares (PLS), Canonical Correlation Analysis (CCA), Support Vector Machine (SVM), and the like, as well as variants thereof.
However, as the industrial modernization process develops, the number of variables collected by the process, the relationships among the variables, the essential characteristics of the variables, and the high-order distribution become more and more complex. And with the development and infiltration of computer technology, a large amount of industrial data is preserved. In such a context, conventional process condition monitoring methods exhibit certain disadvantages, such as: (1) the method comprises the following steps of (1) making a precondition assumption on data not be practical, (2) extracting data features not to be high-order, (3) not being applicable to modeling of mass data, (4) determining the quality of the extracted features by human intervention in the process of extracting the data features, (5) not being robust to detection noise in measurement data, and the like. These problems are extremely disadvantageous for processing industrial big data with complex variable relationships and high-level abstract data features. The results obtained from models constructed in this way also present problems of uncertainty and inaccuracy. Meanwhile, with the development of deep learning technology, the method has unprecedented achievements in the fields of pattern recognition, image classification and the like. As a deep neural network model, the method has the characteristics superior to the traditional method: the method does not need to make any assumption on the acquired data, extract high-order and abstract features, is closer to essential features, has robustness on process measurement noise, and is suitable for mass data.
On the other hand, the intrinsic characteristics of the data measured by the process flow in different states are different, which is not only reflected in the numerical fluctuation of the data and the change of the correlation between the variables (linear correlation, nonlinear correlation, etc.), but also the high-order distribution characteristics of the data show different characteristics in different states. Therefore, a deep learning state monitoring model is constructed, the numerical characteristics of data, the relation characteristics among variables and the change of high-order distribution characteristics can be fully mined in the process of monitoring the real-time working state so as to accurately identify the current working state, and the current working state is effectively judged so as to reasonably further operate the technological process.
Disclosure of Invention
The purpose of the present invention is to utilize all 312 variables in the PTA process flow oxidation section for real-time monitoring of the 4-CBA state space (i.e., the production run state space). 312 global variables such as cobalt concentration (PPM), bromine concentration (PPM), top gas oxygen content (VOL%), oxygen control amount (VOL%), tail oxygen control amount (VOL%), crystallization amount (KNM3/H) and the like are used as input variables of the system and are respectively sent into a Deep auto-encoder (SAE) and a Deep Belief Network (DBN) to construct a Deep learning state monitoring system. Numerical value changes and variable correlation changes in SAE identification data under different process states, high-order distribution feature changes in DBN identification data under different process states, and features extracted under the two depth models are subjected to feature fusion, so that the fused features are full of rich information of the data. And sending the characteristic information into a Softmax classifier to monitor the current working state: an over-oxidized state, a moderately oxidized state, and an under-oxidized state. The state monitoring scheme designed by the invention fully excavates multidirectional characteristic information in data, so that the identification result is more accurate, the related control operation is further guided conveniently, and the production safety is ensured and the product quality is improved.
The invention has the advantages that: (1) a deep neural network is utilized to construct a deep model of the process flow and mine high-order abstract information characteristics in data, so that various problems of the traditional method are solved; (2) numerical characteristics, variable relation characteristics and variable high-order distribution characteristic information in data are fully mined by constructing a double model of SAE and DBN; (3) by fusing the features, the real-time modal identification can be carried out by fully utilizing abundant feature information when the data features are acquired, and the identification precision is improved; (4) constructing a final classifier by using a Softmax classifier, and carrying out modal identification and classification of the process technology in real time; (5) the operating state is divided into an over-oxidation state, a moderate oxidation state and an under-oxidation state.
1. Preprocessing of modeled samples
1.1 outlier detection
In the collected PTA process flow historical data, there may be some "bad" samples and "bad" variables in the applied historical training data due to errors in the process of recording and saving the data. To improve the accuracy of the data applied for modeling, first we perform outlier detection on the data to exclude incomplete samples and inaccurate variables in the data: (1) if there are variables in a sample that are not recorded, then the sample is deleted and not used as training data; (2) if the variable is all 0 in all the recorded samples, the sample has no substantial effect on mining data information and analyzing sample modes, and the variable is deleted and not used as the variable in the training sample.
1.2 normalization Process
In order to eliminate the influence of respective dimensions between variables on the recognition result, the training sample is subjected to standardization preprocessing: each processed dimension data has the characteristics that the mean value is 0 and the variance is 1. For data X epsilon R processed by abnormal valuen×mWhere n is the number of samples and m is the number of variables in a sample, then the preprocessing can be expressed as:
Figure BDA0002624631710000051
wherein x isi∈Rn×1Is the ith sample variable of the training sample X,
Figure BDA0002624631710000052
is the average value of the values of the two,
Figure BDA0002624631710000053
is the variance of the received signal, and,
Figure BDA0002624631710000054
is a normalized sample
Figure BDA0002624631710000055
The ith variable of (1).
2. Modal partitioning
The invention relates to a method for carrying out modal division on the production state of an oxidation section of a PTA process flow based on the content of 4-CBA. In the reaction process, when the content of the 4-CBA is too large, the process flow is in an under-oxidation state; when the 4-CBA content is too small, it means that the process is in a peroxide state. Based on this, the operating state of the oxidation section of the PTA process flow is divided into three states, such as:
Figure BDA0002624631710000061
wherein the content unit of 4-CBA is PPM, 2903PPM and 3001PPM are adjustable control limit values.
3. Depth model construction
3.1 Stack-based self-encoder model construction
A Stacked Autoencoder (SAE) is a neural network with multiple hidden layers. The supervised SAE model used in the present invention can be divided into three parts: input layer, hidden layer, softmax output layer. Given training data
Figure BDA0002624631710000062
And an SAE model with n hidden layers, which is first constructed to obtain the deep features of SAE:
Figure BDA0002624631710000063
wherein [ W ]1,W2,…,Wn]Is the weight of the first hidden layer to the n hidden layer, [ b1,b2,…,bn]Is the threshold value of the first hidden layer to the n hidden layer, [ sigma ]1,σ2,…,σn]Is a nonlinear activation function of the first to the n-th hidden layers, Fsae∈Rn×3Is the high-order feature obtained by the final SAE extraction. As can be seen from the process of extracting the features, the SAE can extract high-order numerical feature information in the data and high-order relation features of linear nonlinearity among the data. This is of great significance to the process of condition monitoring based on data characteristics.
For the ith SAE feature of the original data:
Figure BDA0002624631710000064
the expression after passing through the Softmax classifier can be expressed as:
Figure BDA0002624631710000065
wherein p (y)1),p(y2),p(y3) Are respectively a sample fsae iThe probability of the first class, the second class and the third class, wherein the class corresponding to the maximum probability is used as the final fsae iThe category to which it belongs. And adjusting the parameters of the network by adopting an optimization algorithm according to the error between the prediction category and the real category until the network is stable.
3.2 deep belief network model construction
A Deep Belief Network (DBN) is a deep neural network formed by a plurality of Restricted Boltzmann Machines (RBMs) superimposed on one another. The RBM is a neural network (a visible layer v and a hidden layer h) with two layers of neurons, the two layers are symmetrically connected in a bidirectional way, and the neurons in each layer are independent. RBM is a probability model based on an energy function, where the energy function can be expressed as:
Figure BDA0002624631710000071
wherein
Figure BDA0002624631710000072
bi,cjThreshold values of apparent and hidden layers, wijIs the weight connecting the explicit and implicit layers. But the neurons in a conventional RBM are binary, i.e., the value of a neuron is only 0 or 1. However, the variables collected in the process flow are continuous values, so the DBN used in the present invention is composed of gaussian rbm (grbm). GRBM allows the input of continuous values whose energy function can be expressed as:
Figure BDA0002624631710000073
wherein sigmai 2Is the variance of the gaussian noise added to the neuron. Based on the energy function, the training of the RBM employs a contrast-divergence (CD-k) algorithm for sampling to extract distribution features in the data. The DBN is formed by overlapping a plurality of RBMs, and the hidden layer of the previous RBM is the display layer of the next RBM. In this way, the DBN can extract a high-order distribution feature, F, in the datadbn∈Rn ×3. Likewise, the features are classified and trained using a Softmax classifier.
4. Feature fusion and classifier construction
For the depth features obtained by SAE and DBN training, they are feature fused. The fused features can be expressed as:
F＝[Fdbn；Fsae] (7)
and training a Softmax classifier based on the fused features to classify the data according to the rich data features.
5. Production state monitoring technology based on SAE/DBN-Softmax model
The off-line modeling and on-line monitoring process of the oxidation section production running state space monitoring method in the PTA process flow is shown in FIG. 2. Historical data was first collected and recorded, and then outlier detection was performed on the training data to remove "bad" samples and "bad" variables, as described in section 1.1. The data was then normalized and the mean and variance of the training data were recorded as described in section 1.2. And then respectively sending the modeling data into SAE and DBN for supervised modeling, classifying the samples by utilizing Softmax, namely, monitoring the state of the process, and optimizing respective network according to the error between the predicted value and the true value until the network is stable. As described in sections 3.1 and 3.2. The features extracted from SAE and DBN are then feature fused as described in section 4. And finally, sending the data to a Softmax classifier for final recognition and model adjustment. When online identification is carried out, the data monitored in real time are standardized by using the mean value and the variance of training data, and then the data are sent to the constructed SAE model and the DBN model in real time for feature extraction. And fusing the extracted features, and sending the fused features into a trained Softmax classifier for state monitoring.
Drawings
FIG. 1 is a block diagram of a PTA production process.
FIG. 2 is a PTA process flow 4-CBA status monitoring flow diagram.
Detailed Description
The invention is further illustrated by the following examples:
for 312 global variables of the PTA process flow oxidation section, 340 samples were first collected to construct training data, and 340 samples were modal classified according to the 4-CBA content (according to formula 2), wherein 130 samples were moderately oxidized, 85 samples were peroxidized, and 125 samples were underoxidized. And a set of 40 samples of the data of the implementation test was prepared, 15 moderately oxidized samples, 10 excessively oxidized samples, and 15 insufficiently oxidized samples.
1. Pretreating samples
The above collected training data is subjected to outlier detection (e.g., 1.1) and normalization of the data (e.g., 1.2). In the abnormal value detection process, a case that a certain variable record of a sample is missing exists, and therefore the sample is deleted. Then, normalization is performed by first calculating the mean and variance of the 312-dimensional variables, where X is ═ X for the training sample variables1,x2,x3,…,x311,x312]And (4) showing. By calculation, x1,x2,x3,…,x311,x312The mean value of (A) is: 237.593,360.076,330.033, …,149.41, 79.0755; the variance is: 12599.8,32678.6,43074.5, …,341.589, 58.735. A normalization calculation is then performed:
Figure BDA0002624631710000091
Figure BDA0002624631710000092
Figure BDA0002624631710000093
…
Figure BDA0002624631710000094
Figure BDA0002624631710000101
for data x acquired in real timetThe same mean and variance were used for normalization:
Figure BDA0002624631710000102
wherein, the mean vector μ ═ 237.593,360.076,330.033, …,149.41,79.0755], and the variance vector σ ═ 2599.8,32678.6,43074.5, …,341.589, 58.735.
2. Recognition model construction based on SAE/DBN-Softmax
And carrying out preliminary modeling on training data by using SAE and DBN, and constructing a Softmax state monitoring and recognition classifier by using the fused features after extracting the data features. The specific model parameters are as follows:
(1) firstly, performing SAE modeling on the preprocessed data according to 3.1, wherein the structure adopted by the SAE model is 312-:
Fsae＝[0.494327,-0.805896,-0.350029,-0.981692,0.999401,0.999977,-0.923577,
0.793337,0.964532,0.999917,0.990558,0.999802,-0.754223,-0.787286,-0.249494,
-0.937339,-0.969371,...
...
-0.905926,0.395921,-0.0161926,0.990046,0.786132,0.888272,0.938116,-0.996132,
-0.983115,0.999472,0.266457,-0.531727,-0.911818,0.981601,-0.999524,-0.999249,
-0.989958；
...]∈R339×250
only 250 higher-order SAE features of the first set of training data are listed here. The resulting SAE training classification based on these 250 features is:
Ssae＝[1 3.62333e-10 9.67149e-11
1 4.36183e-09 4.40108e-09
1 9.2236e-12 8.7458e-10
...
4.86728e-09 1 1.36423e-14
1.02999e-09 1 9.71407e-15
7.79071e-10 1 1.51517e-13
...
6.02575e-10 1.03221e-15 1
2.0055e-08 3.19144e-14 1
4.7929e-10 2.06275e-15 1
...]∈R339×3
(2) and then according to 3.2, carrying out DBN modeling on the preprocessed data, wherein the structure adopted by the DBN is 312-:
Fdbn＝[0.886368,-0.870236,-0.939022,-0.882945,0.935839,0.999422,
0.474037,-0.99793,0.809049,0.99972,0.982737,-0.958993,-0.459232,
0.656874,0.999083,-0.599752,-0.959192,0.537159,0.260626,0.918603
...
-0.996811,0.353469,-0.725276,0.98191,-0.497436,0.454034,-0.182224,
0.872963,0.870272,-0.986479,-0.98474,0.994634,0.623986,-0.858509,
0.957397,-0.283898,-0.936901,-0.999997,-0.97561；
...]∈R339×250
only the DBN high-order distribution characteristics of the first group of data are listed, and the classification result of the DBN can be obtained by performing Softmax through the characteristics:
Sdbn＝[1 3.36395e-10 1.97449e-11
1 3.97148e-10 2.67883e-10
1 2.08636e-09 2.48129e-10
...
1.46696e-10 1 3.15919e-11
1.27659e-08 1 1.00844e-08
1.0514e-08 1 1.52888e-12
...
1.03352e-08 8.85584e-14 1
3.0354e-09 9.37431e-14 1
6.43972e-09 1.5605e-12 1
...]∈R339×3
(3) f is to besaeAnd FdbnCarrying out fusion of F ═ Fsae；Fdbn]And sent into a Softmax classifier, wherein the structure of the classifier is 500-3. And adjusting the parameters of the final Softmax classifier based on the fused features.
And (3) carrying out state monitoring and recognition on the following test data samples collected in real time by using the recognition model obtained by the training data:
[x1,x2,x3,...,x311,x312]T＝[305.694,453.11,464.185,...,159.071,80.0502]T
after pretreatment, the following are obtained:
Figure BDA0002624631710000121
the characteristics obtained by the constructed SAE and DBN models are as follows:
fsae+dbn＝[-0.899105,0.488153,0.978834,-0.989181,-0.352779,0.647489,
-0.983002,-0.943049,-0.999096,0.715804,-0.957282,-0.993672,
0.976676,0.971916,0.408356,-0.990261,-0.99817,0.756899,
0.00159877,0.673206
...
0.434951,0.606158,0.931484,-0.999018,-0.1693,-0.520611,
-0.986429,-0.215725,-0.980154,0.960388,-0.0828553,-0.999595,
0.992475,-0.975431,0.752204,-0.993587,-0.999844,-0.925812,
0.997847,0.999205]∈R1×500
sending the fusion characteristics into a Softmax classifier, and finally obtaining a result as follows:
s＝[1.05454e-05 0.99999 5.89481e-13]
the results suggest that there is 99.999% reason to believe that the process flow for this sample is under-oxidized.

Claims (6)

  1. The method is characterized in that a working state monitoring model is established by taking a global variable in the process flow as an input variable so as to identify the working state of the industrial process at the current moment;
    the global variable is a state space detection of the content of the byproduct 4-CBA in the production process so as to reflect the working state of the current process flow;
    the working state monitoring model is established by constructing a deep neural network state monitoring model to extract high-order abstract characteristic information in data;
    extracting high-order numerical value feature information in the data and high-order correlation features among variables by using a depth self-encoder SAE, and extracting high-order distribution feature information in the data by using a depth confidence network DBN;
    features of multiple scales are fused to enrich the expression of many aspects of the data. And (3) integrating various characteristics of the data, and monitoring and identifying the state by using a Softmax classifier.
  2. 2. The method for spatially monitoring the operating conditions of an oxidation section in a PTA process according to claim 1, wherein said establishing a model for monitoring the operating conditions comprises the steps of:
    preprocessing of modeled samples
    1.1 outlier detection
    In the collected PTA process flow historical data, because errors exist in the process of recording and storing the data, some 'bad' samples and 'bad' variables exist in the applied historical training data; to improve the accuracy of the data applied for modeling, first we perform outlier detection on the data to exclude incomplete samples and inaccurate variables in the data: (1) if there are variables in a sample that are not recorded, then the sample is deleted and not used as training data; (2) if the variable is all 0 in all the record samples, the sample has no substantial effect on mining data information and analyzing sample modes, and the variable is deleted and not used as the variable in the training sample;
    1.2 normalization Process
    In order to eliminate the influence of respective dimensions between variables on the recognition result, the training sample is subjected to standardization preprocessing: enabling each processed dimension data to have the characteristics of 0 mean value and 1 variance; for data X epsilon R processed by abnormal valuen ×mWhere n is the number of samples and m is the number of variables in a sample, then the preprocessing can be expressed as:
    Figure FDA0002624631700000021
    wherein x isi∈Rn×1Is the ith sample variable of the training sample X,
    Figure FDA0002624631700000022
    is the average value of the values of the two,
    Figure FDA0002624631700000023
    is the variance of the received signal, and,
    Figure FDA0002624631700000024
    is a normalized sample
    Figure FDA0002624631700000025
    The ith variable of (1).
  3. 3. The method for spatially monitoring the operating conditions of an oxidation section in a PTA process according to claim 1, wherein said establishing a model for monitoring the operating conditions comprises the steps of:
    modal partitioning
    Carrying out modal division on the production state of an oxidation section of a PTA process flow based on the content of 4-CBA: in the reaction process, when the content of the 4-CBA is too large, the process flow is in an under-oxidation state; when the 4-CBA content is too small, it means that the process is in a peroxide state. Based on this, the operating state of the oxidation section of the PTA process flow is divided into three states, such as:
    Figure FDA0002624631700000026
    wherein the content unit of 4-CBA is PPM, 2903PPM and 3001PPM are adjustable control limit values.
  4. 4. The method for spatially monitoring the operating conditions of an oxidation section in a PTA process according to claim 1, wherein said establishing a model for monitoring the operating conditions comprises the steps of:
    depth model construction
    3.1 Stack-based self-encoder model construction
    The stacked self-encoder SAE is divided into three parts: an input layer, a hidden layer and a softmax output layer; given training data
    Figure FDA0002624631700000031
    And an SAE model with n hidden layers, which is first constructed to obtain the deep features of SAE:
    Figure FDA0002624631700000032
    wherein [ W ]1,W2,…,Wn]Is the weight of the first hidden layer to the n hidden layer, [ b1,b2,…,bn]Is the threshold value of the first hidden layer to the n hidden layer, [ sigma ]1,σ2,…,σn]Is a nonlinear activation function of the first to the n-th hidden layers, Fsae∈Rn×3Is the high-order feature obtained by the final SAE extraction. Can be extracted fromIn the characterization process, SAE can extract high-order numerical characteristic information in data and linear nonlinear high-order relation characteristics among data;
    for the ith SAE feature of the original data:
    Figure FDA0002624631700000033
    the expression after passing through the Softmax classifier can be expressed as:
    Figure FDA0002624631700000034
    wherein p (y)1),p(y2),p(y3) Are respectively a sample fsae iThe probability of the first class, the second class and the third class, wherein the class corresponding to the maximum probability is used as the final fsae iThe category to which it belongs; adjusting and optimizing the parameters of the network by adopting an optimization algorithm according to the error between the prediction category and the real category until the network is stable;
    3.2 deep belief network model construction
    A Deep Belief Network (DBN) is a deep neural network formed by a plurality of Restricted Boltzmann Machines (RBMs) superimposed on one another. The RBM is a neural network (a visible layer v and a hidden layer h) with two layers of neurons, the two layers are symmetrically connected in a bidirectional way, and the neurons in each layer are independent. RBM is a probability model based on an energy function, where the energy function can be expressed as:
    Figure FDA0002624631700000041
    wherein
    Figure FDA0002624631700000042
    bi,cjThreshold values of apparent and hidden layers, wijIs the weight connecting the apparent layer and the hidden layer; the applied DBN is composed of Gaussian RBM (GRBM); GRBM allowed continuous numberThe input of a value, its energy function can be expressed as:
    Figure FDA0002624631700000043
    wherein sigmai 2Is the variance of the gaussian noise added to the neuron; based on an energy function, the RBM training adopts a contrast-divergence (CD-k) algorithm to sample so as to extract distribution characteristics in data; the DBN is formed by overlapping a plurality of RBMs, and the hidden layer of the previous RBM is the display layer of the next RBM; in this way, the DBN can extract a high-order distribution feature, F, in the datadbn∈Rn×3(ii) a The features are also classified and trained using a Softmax classifier.
  5. 5. The method for spatially monitoring the operating conditions of an oxidation section in a PTA process according to claim 1, wherein said establishing a model for monitoring the operating conditions comprises the steps of:
    feature fusion and classifier construction
    Performing feature fusion on the deep features obtained by SAE and DBN training; the fused features can be expressed as:
    F＝[Fdbn；Fsae] (7)
    and training a Softmax classifier based on the fused features to classify the data according to the rich data features.
  6. 6. The method for spatially monitoring the operating conditions of an oxidation section in a PTA process according to claim 1, wherein said establishing a model for monitoring the operating conditions comprises the steps of:
    production state monitoring based on SAE/DBN-Softmax model
    The method for monitoring the production running state space of the oxidation working section in the PTA process flow comprises the following steps of off-line modeling and on-line monitoring: firstly, collecting and recording historical data, and then carrying out abnormal value detection on training data to remove 'bad' samples and 'bad' variables; then, carrying out standardization processing on the data, and recording the mean value and the variance of the training data; then, respectively sending modeling data into SAE and DBN for supervised modeling, classifying samples by utilizing Softmax, namely, monitoring the state of the process, and optimizing respective network according to the error between a predicted value and a true value until the network is stable; then, performing feature fusion on features extracted by SAE and DBN; finally, sending the data to a Softmax classifier for final recognition and model adjustment; when online identification is carried out, carrying out standardization processing on data monitored in real time by using the mean value and the variance of training data, and then sending the data into a constructed SAE model and a DBN model in real time for feature extraction; and fusing the extracted features, and sending the fused features into a trained Softmax classifier for state monitoring.
CN202010793556.6A 2020-08-10 2020-08-10 Method for monitoring production running state space of oxidation section in PTA (pure terephthalic acid) process flow Active CN112034793B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010793556.6A CN112034793B (en) 2020-08-10 2020-08-10 Method for monitoring production running state space of oxidation section in PTA (pure terephthalic acid) process flow

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010793556.6A CN112034793B (en) 2020-08-10 2020-08-10 Method for monitoring production running state space of oxidation section in PTA (pure terephthalic acid) process flow

Publications (2)

Publication Number Publication Date
CN112034793A true CN112034793A (en) 2020-12-04
CN112034793B CN112034793B (en) 2024-04-26

Family

ID=73576791

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010793556.6A Active CN112034793B (en) 2020-08-10 2020-08-10 Method for monitoring production running state space of oxidation section in PTA (pure terephthalic acid) process flow

Country Status (1)

Country Link
CN (1) CN112034793B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112904810A (en) * 2021-01-13 2021-06-04 中南大学 Process industry nonlinear process monitoring method based on effective feature selection
CN113091808A (en) * 2021-03-29 2021-07-09 江苏利宏科技发展有限公司 Chemical industry industrial control instrument with comprehensive information management system and system thereof
CN115755823A (en) * 2023-01-10 2023-03-07 西北电子装备技术研究所(中国电子科技集团公司第二研究所) Electrical control system applied to eutectic surface mounting equipment

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106555788A (en) * 2016-11-11 2017-04-05 河北工业大学 Application of the deep learning based on Fuzzy Processing in hydraulic equipment fault diagnosis
CN107885914A (en) * 2017-10-25 2018-04-06 华东理工大学 Industrial reactor running status visualizes on-line monitoring method
CN110739031A (en) * 2019-11-18 2020-01-31 中南大学 Method and device for supervised prediction of metallurgical sintering processes and storage medium

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106555788A (en) * 2016-11-11 2017-04-05 河北工业大学 Application of the deep learning based on Fuzzy Processing in hydraulic equipment fault diagnosis
CN107885914A (en) * 2017-10-25 2018-04-06 华东理工大学 Industrial reactor running status visualizes on-line monitoring method
CN110739031A (en) * 2019-11-18 2020-01-31 中南大学 Method and device for supervised prediction of metallurgical sintering processes and storage medium

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
JIANBO YU，ET AL.: "Active features extracted by deep belief network for process monitoring", ISA TRANSACTIONS, vol. 84, no. 2019, pages 247 - 261 *
刘瑞兰 等: "模糊神经网络的混合学习算法及其软测量建模", 系统仿真学报, vol. 17, no. 12, pages 2878 - 2881 *
张祥 等: "基于 VAE-DBN 的故障分类方法在化工过程中的应用", 过程工程学报, vol. 18, no. 3, pages 590 - 594 *
高国正 等: "洛阳PTA装置基于径基函数网络的4-CBA软测量", 聚酯工业, vol. 18, no. 6, pages 11 - 14 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112904810A (en) * 2021-01-13 2021-06-04 中南大学 Process industry nonlinear process monitoring method based on effective feature selection
CN112904810B (en) * 2021-01-13 2022-07-01 中南大学 Process industry nonlinear process monitoring method based on effective feature selection
CN113091808A (en) * 2021-03-29 2021-07-09 江苏利宏科技发展有限公司 Chemical industry industrial control instrument with comprehensive information management system and system thereof
CN115755823A (en) * 2023-01-10 2023-03-07 西北电子装备技术研究所(中国电子科技集团公司第二研究所) Electrical control system applied to eutectic surface mounting equipment

Also Published As

Publication number Publication date
CN112034793B (en) 2024-04-26

Similar Documents

Publication Publication Date Title
CN112034793B (en) Method for monitoring production running state space of oxidation section in PTA (pure terephthalic acid) process flow
CN112001527B (en) Industrial production process target data prediction method of multi-feature fusion depth neural network
JP5301310B2 (en) Anomaly detection method and anomaly detection system
Tian et al. Identification of abnormal conditions in high-dimensional chemical process based on feature selection and deep learning
Liu et al. Toward robust fault identification of complex industrial processes using stacked sparse-denoising autoencoder with softmax classifier
CN112198842B (en) Tracing diagnosis method for abnormal operation state of PTA oxidation unit
CN112363462B (en) Static-dynamic cooperative sensing complex industrial process running state evaluation method
Cao et al. Multimodal process monitoring based on variational Bayesian PCA and Kullback-Leibler divergence between mixture models
CN117542169A (en) Automatic equipment temperature abnormality early warning method based on big data analysis
Zhang et al. Amplitude‐frequency images‐based ConvNet: Applications of fault detection and diagnosis in chemical processes
CN111913447B (en) Online data-driven nonlinear chemical process monitoring method
CN107121490B (en) The evaluation method of operating status in a kind of hydrometallurgy process
WO2022058116A1 (en) Explaining machine learning output in industrial applications
Zhou et al. Online abnormal interval detection and classification of industrial time series data based on multi-scale deep learning
Chang et al. Multimode process mode identification with coexistence of quantitative information and qualitative information
CN116714437B (en) Hydrogen fuel cell automobile safety monitoring system and monitoring method based on big data
CN111914888A (en) Chemical process monitoring method integrating multi-working-condition identification and fault detection
CN112799382A (en) Robot micro-electro-mechanical system fault diagnosis method and system
CN116224887A (en) Multi-station intermittent process fault monitoring method
CN114527731A (en) Industrial process running state evaluation method based on supervision probability slow characteristic analysis
CN111914886B (en) Nonlinear chemical process monitoring method based on online brief kernel learning
CN113551904A (en) Gear box multi-type concurrent fault diagnosis method based on hierarchical machine learning
CN111914214A (en) PTA production process soft measurement method based on reduced KPLS model
SOUZA et al. Trend-weighted rule-based expert system with application to industrial process monitoring
CN114024820B (en) Equipment health portrait method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant